ASTRA: Mapping Art-Technology Institutions via Conceptual Axes, Text Embeddings, and Unsupervised Clustering

Joonhyung Bae

cs.DLMar 28, 2026

ASTRA: Mapping Art-Technology Institutions via Conceptual Axes, Text Embeddings, and Unsupervised Clustering

Authors: Joonhyung Bae

Organizations: Graduate School of Culture Technology, KAIST (Korea Advanced Institute of Science and Technology), Daejeon, South Korea

Abstract

The global landscape of art-technology institutions, including festivals, biennials, research labs, conferences, and hybrid organizations, has grown increasingly diverse, yet systematic frameworks for analyzing their multidimensional characteristics remain scarce. This paper proposes ASTRA (Art-technology Institution Spatial Taxonomy and Relational Analysis), a computational methodology combining an eight-axis conceptual framework (Curatorial Philosophy, Territorial Relation, Knowledge Production Mode, Institutional Genealogy, Temporal Orientation, Ecosystem Function, Audience Relation, and Disciplinary Positioning) with a text-embedding and clustering pipeline to map 78 cultural-technology institutions into a unified analytical space. Each institution is characterized through qualitative descriptions along the eight axes, encoded via E5-large-v2 sentence embeddings and quantized through a word-level codebook into TF-IDF feature vectors. Dimensionality reduction using UMAP, followed by agglomerative clustering (Average linkage, k=10), yields a composite score of 0.825, a silhouette coefficient of 0.803, and a Calinski-Harabasz index of 11196. Non-negative matrix factorization extracts ten latent topics, and a neighbor-cluster entropy measure identifies boundary institutions bridging multiple thematic communities. An interactive React-based tool enables curators, researchers, and policymakers to explore institutional similarities and cross-disciplinary connections. Results reveal coherent groupings such as an art-science hub cluster anchored by ZKM and ArtScience Museum, an innovation and industry cluster including Ars Electronica, transmediale, and Sonar, an ACM academic cluster comprising TEI, DIS, and NIME, and an electronic music cluster including CTM Festival, MUTEK, and Sonic Acts. Code and data: https://github.com/joonhyungbae/astra

Figures & tables

Study	Unit	Representation	Clustering	Topics	Viz Tool	Domain
Manovich [ 15 ]	Artifacts	Image/text features	—	—	Custom	General culture
Stoltz & Taylor [ 24 ]	Media corpora	Word2vec	—	—	—	Immigration disc.
Wang & Goh [ 29 ]	Game reviews	BoW / TF-IDF	—	LDA	—	Entertainment
Sim et al. [ 23 ]	Game reviews	BoW / TF-IDF	—	LDA	—	Entertainment
Yetisen et al. [ 32 ]	Survey	Metadata catalog	—	—	Web catalog	Bioart
Bae [ 3 ]	Artworks	Dual LLM emb. + codebook	Agglom.	—	React	Bioart

Table 1: Comparative summary of closely related studies and the present work. Columns indicate the unit of analysis, the text-representation method, the clustering approach, topic modeling, interactive visualization, and the target domain.

Panel of Figure 1 — Figure 1: Overview of the ASTRA processing pipeline. Qualitative axis descriptions for 78 institutions are encoded through E5-large-v2 sentence embeddings, quantized via a word-level codebook, and clustered using UMAP and agglomerative clustering. NMF topic modeling and entropy-based boundary analysis are applied post-clustering, and results are served through an interactive web visualization.

Axis	Core Question	Theoretical Basis
Curatorial Philosophy	What does the institution stand for?	Hoetzlein; Bourdieu
Territorial Relation	How does it relate to place?	Quinn; Négrier
Knowledge Production	What knowledge does it produce and how?	Hoetzlein; UNESCO FCS
Institutional Genealogy	Where does it come from?	DiMaggio & Powell
Temporal Orientation	How does it relate to time?	Grau
Ecosystem Function	What role does it play in the ecosystem?	Hannan & Freeman [ 12 ]

Table 2: Mapping of the eight conceptual axes to their theoretical foundations.

Algorithm	$k$	Composite	Silhouette	CH	DB	Noise%
Agglomerative (Average)	10	0.825	0.803	11,196	0.593	0.0%
OPTICS	10	0.818	0.931	18,934	0.348	18.75%
$k$ -means	13	0.813	0.702	10,326	0.508	0.0%
DBSCAN	2	—	0.998	—	—	27.5%

Table 3: Comparison of clustering algorithms (best configuration per algorithm).

Figure 2: 2D UMAP scatter plots comparing Agglomerative Average ( $k{=}10$ , left) and DBSCAN ( $k{=}2$ , right). Agglomerative clustering yields 10 interpretable groups, while DBSCAN produces only two effective clusters with 27.5% of institutions classified as noise (gray).

Configuration	Silhouette	CH Index	DB Index
E5-large-v2 (1,024-d, $k{=}5$ ) [ 28 ]	0.845	8,752	0.494
GTE-Qwen2+SBERT dual ( $k{=}5$ )	0.836	10,171	0.503
GTE-Qwen2+BGE dual ( $k{=}7$ )	0.831	13,962	0.556
SBERT-Mini+BGE dual ( $k{=}7$ )	0.828	7,451	0.526
E5-large-v2 ( $k{=}7$ )	0.825	14,317	0.556
Nomic ( $k{=}9$ )	0.786	14,171	0.589

Table 4: Embedding comparison (sweep-optimized: best codebook

k

, linkage, and cluster count per config; metrics computed in post-UMAP space). Parenthesized

k

denotes codebook size. Values differ from Table 3 (which fixes cluster

k{=}10

, Average linkage) because each row reports its own sweep-optimal cluster count (ranging from

k{=}8

k{=}13

Removed Axis	Sil.	$\Delta_{\text{sil}}$	CH	$\Delta_{\text{CH}}$
(none — baseline)	0.399	—	49.70	—
Curatorial Philosophy	0.366	+0.033	55.38	$-5.67$
Territorial Relation	0.395	+0.004	48.17	+1.54
Knowledge Production Mode	0.418	$-0.019$	62.91	$-13.20$
Institutional Genealogy	0.352	+0.047	43.88	+5.83
Temporal Orientation	0.416	$-0.017$	51.48	$-1.78$

Table 5: Leave-one-axis-out analysis.

\Delta_{\text{sil}}

and

\Delta_{\text{CH}}

: changes relative to baseline (positive

\Delta_{\text{sil}}

= removing the axis decreases quality).

Figure 3: Axis contribution analysis (leave-one-axis-out). Each bar represents the change in the respective metric when the named axis is removed.

Figure 4: 2D UMAP scatter plot of 78 institutions colored by cluster membership ( $k{=}10$ ). (a) Full view showing the spatial separation of Cluster 4; (b) zoomed view of the nine remaining clusters with representative institution labels.

Cl.	$n$	Years	Dominant Topic	Representative Institutions
0	6	1969–2014	Computer Art & Tech	BIAN, ACM SIGGRAPH, V2_ Lab, e-flux, NEW INC
1	7	1974–2015	Innovation & Industry	Ars Electronica, transmediale, Sónar, SXSW, ICMC
2	8	1837–2012	Creative Community	Eyeo, FILE, Resonate, EVA, ITP NYU
3	9	1990–2013	Biennials & Confer.	NODE Forum, SeMA Mediacity Seoul, Biennale Némo, ISEA
4	24	1970–2017	Mixed (Large)	Vivid Sydney, Piksel, SMC, Lumen Prize, Japan Media Arts
5	7	1989–2011	Music & Sound Art	CTM, MUTEK, Elektra, Sonic Acts, ACM IMX

Table 6: Summary of 10 clusters: size, founding year range, dominant topic, and representative institutions.

Figure 5: Cluster–topic heatmap. Each cell shows the mean NMF weight of the topic (column) within the cluster (row). Darker cells indicate stronger thematic associations.

Institution	Cl.	Type	Top-3 Neighbor Clusters
Ars Electronica Festival	1	Festival	5, 1, 2
ZKM \| Center for Art and Media	9	Center	9, 4, 8
EMPAC	8	Residency	8, 4, 9
Chronus Art Center (CAC)	8	Center	8, 9
ACM TEI	7	Conference	7, 3
SIGGRAPH Asia	9	Conference	9, 4

Table 7: Selected boundary institutions with high neighbor-cluster entropy.

Figure 6: Screenshot of the APESuite Explorer web interface, showing the 2D scatter plot, selected institution detail panel with thematic profile, and similarity links.

M-ArtAgent: Evidence-Based Multimodal Agent for Implicit Art Influence Discovery
Apr 8, 2026Hanyi Liu, Zhonghao Jiu, Minghao Wang +2Multimodal Agents Multimodal Reasoning
China Electronics Technology Group Co., Ltd., Beijing 100043, China · School of Information Science and Engineering, Southeast University, Nanjing 211189, China · Department of Chemical and Biological Engineering, Hong Kong University of Science and Technology, Hong Kong SAR, China +2
Iconographic Classification and Content-Based Recommendation for Digitized Artworks
Feb 23, 2026Krzysztof Kutt, Maciej BaczyńskiDigital Cultural Heritage Image Retrieval
Department of Human-Centered Artificial Intelligence, Institute of Applied Computer Science, Faculty of Physics, Astronomy and Applied Computer Science, Jagiellonian University, prof. Stanisława Łojasiewicza 11, 30-348 Kraków, Poland
ArtSeek: Deep artwork understanding via multimodal in-context reasoning and late interaction retrieval
Jul 29, 2025Nicola Fanelli, Gennaro Vessio, Giovanna CastellanoMultimodal RAG Multimodal IR
Department of Computer Science, University of Bari Aldo Moro, Bari, Italy

Apr 8, 2026J/K move · Enter open · S save

M-ArtAgent: Evidence-Based Multimodal Agent for Implicit Art Influence Discovery

Hanyi Liu, Zhonghao Jiu, Minghao Wang, Yuhang Xie+1

China Electronics Technology Group Co., Ltd., Beijing 100043, China · School of Information Science and Engineering, Southeast University, Nanjing 211189, China · Department of Chemical and Biological Engineering, Hong Kong University of Science and Technology, Hong Kong SAR, China+2

Multimodal Agents Multimodal Reasoning

Open paper AskPDF arXiv

Implicit artistic influence, although visually plausible, is often undocumented and thus poses a historically constrained attribution problem: resemblance is necessary but not sufficient evidence. Most prior systems reduce influence discovery to embedding similarity or label-driven graph completion, while recent multimodal large language models (LLMs) remain vulnerable to temporal inconsistency and unverified attributions. This paper introduces M-ArtAgent, an evidence-based multimodal agent that reframes implicit influence discovery as probabilistic adjudication. It follows a four-phase protocol consisting of Investigation, Corroboration, Falsification, and Verdict governed by a Reasoning and Acting (ReAct)-style controller that assembles verifiable evidence chains from images and biographies, enforces art-historical axioms, and subjects each hypothesis to adversarial falsification via a prompt-isolated critic. Two theory-grounded operators, StyleComparator for Wolfflin formal analysis and ConceptRetriever for ICONCLASS-based iconographic grounding, ensure that intermediate claims are formally auditable. On the balanced WikiArt Influence Benchmark-100 (WIB-100) of 100 artists and 2,000 directed pairs, M-ArtAgent achieves 83.7% positive-class F1, 0.666 Matthews correlation coefficient (MCC), and 0.910 area under the receiver operating characteristic curve (ROC-AUC), with leakage-control and robustness checks confirming that the gains persist when explicit influence phrases are masked. By coupling multimodal perception with domain-constrained falsification, M-ArtAgent demonstrates that implicit influence analysis benefits from historically grounded adjudication rather than pattern matching alone.

Abstract

Explore similar work

M-ArtAgent: Evidence-Based Multimodal Agent for Implicit Art Influence Discovery

Iconographic Classification and Content-Based Recommendation for Digitized Artworks

ArtSeek: Deep artwork understanding via multimodal in-context reasoning and late interaction retrieval